Google DeepMind Frontier Safety Framework

This document sets out Google DeepMind's internal rules for identifying and managing the most serious risks from its AI models before those models are widely released. No model can be broadly deployed until a designated internal governance body reviews and approves a documented safety case. Google DeepMind also acknowledges that some of its current safety monitoring tools are still being developed and that its security measures deliver their full benefit only if adopted broadly across the AI industry.

Google DeepMind's Frontier Safety Framework establishes a structured process for assessing and mitigating risks associated with critical capability levels in its AI models. The Framework mandates that all critical capabilities undergo a thorough mitigation process beginning with iterative safeguard preparation, culminating in a safety case—an assessable argument demonstrating that severe risks have been minimised to an acceptable level—which must receive affirmative approval from an appropriate corporate governance body before any general availability deployment. Security mitigations are framed as foundational, with model weight protection identified as essential because exfiltration of weights enables removal of most safeguards; the Framework additionally recommends particularly high security levels for critical capabilities in the machine learning R&D domain. Google DeepMind acknowledges that its automated monitoring of instrumental reasoning capabilities is exploratory rather than fully implemented, and explicitly recognizes that automated monitoring is not expected to remain sufficient long-term if models reach stronger levels of instrumental reasoning. The Framework further establishes a qualified commitment to share information with appropriate government authorities when Google DeepMind assesses that a model presents an unmitigated and material risk to overall public safety.

For an individual user, this document means that Google DeepMind has committed to clearing a mandatory internal approval gate before broadly releasing any model assessed as carrying critical capability risks. Security protections specifically cover model weights, because their loss would undermine most other safeguards that protect users. The Framework's disclosure commitment—sharing information with government authorities when an unmitigated and material public-safety risk is assessed—is qualified by Google DeepMind's own assessment and framed as an aim rather than an unconditional obligation, which users should weigh when considering the reliability of that protection.